PRISM -- Thin-Film Optical Design Dataset

Project Page | Paper | Github

Synthetic dataset of thin-film multilayer optical stacks and their simulated reflectance/transmittance spectra. Generated for training and evaluating PRISM (Position-encoded Regressive Inverse Spectral Model), an autoregressive transformer for inverse thin-film design.

Each sample is a (materials, thicknesses, spectrum) triple: a thin-film stack definition and its physically simulated optical response.

Subsets

Training data

Subset	Layers	Thickness step	Thickness range	Splits	Total samples
max_len_20_10nm	1--20	10 nm	10--500 nm	train / dev / val	10,110,000

Validation-only (out-of-distribution)

These subsets have no training split and are used to evaluate generalisation.

Subset	Layers	Thickness step	Thickness range	Samples	Purpose
max_len_20_5nm	1--20	5 nm	5--250 nm	110,000	Dev + val for 5 nm (alternate path)
max_len_20_15nm	1--20	15 nm	15--750 nm	20,000	OOD thickness step
max_len_20_20nm	1--20	20 nm	20--1000 nm	10,000	OOD thickness step
max_len_30_10nm	20--30	10 nm	10--500 nm	10,000	OOD sequence length
max_len_40_10nm	30--40	10 nm	10--500 nm	10,000	OOD sequence length
`max_len_50_10nm"	40--50	10 nm	10--500 nm	10,000	OOD sequence length
thick/15nm	20	15 nm	15--750 nm	10,000	Thick designs only (cum. depth >= 11,000 nm)
thick/20nm	20	20 nm	20--1000 nm	10,000	Thick designs only (cum. depth >= 11,000 nm)
max_length_10	1--10	5 nm	5--250 nm	30,000	Short sequence validation

Citation

@misc{wang2024prism,
      title={PRISM: Position-encoded Regressive Inverse Spectral Model for Multilayer Thin-Film Design}, 
      author={Runtian Wang and Renhao Xue and Baige Chen and Hao Wu},
      year={2024},
      eprint={2605.26502},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}